Part 4 · Using software and AI · Chapter 4.14
AI reading documents
Draft: the wording may change.
The idea
AI can read a photo of a bill, a handwritten order or a bank slip and fill in the entry for you. What it produces is a proposed entry — a draft waiting for a person — never a posted one. The photo stays attached as the evidence, so whoever approves can look at both side by side.
From photo to proposed entry
The steps are the same whatever the paper is:
- Capture. Someone photographs the document, flat, in good light, the whole page in frame.
- Read. The AI picks out the fields: party, bill number, date, items, quantities, rates, tax, total.
- Match. The software looks up what it read against your masters: which supplier, which product, which tax rate.
- Propose. It shows a draft entry, with the photo beside it and any doubtful field marked.
- Approve. A person checks, corrects and posts. Only now does it touch the books.
Steps 1 to 4 save the typing. Step 5 is where responsibility sits, and it is never skipped.
At the counter
A company invoice arrives with the morning's delivery: 14 lines, total ₹1,18,000 including GST. Typing it in takes the clerk about 20 minutes. Photographed, the draft is ready in under a minute:
| What AI read | What the clerk found |
|---|---|
| Supplier, invoice number, date | Correct |
| 13 of 14 product lines | Correct, matched to the product master |
| Line 9: "Biscuit 75g × 40" | The invoice says 48 — the 8 was folded under a crease |
| Tax and total | Total matched the paper, because AI copied the printed total, not its own sum |
That last line is the trap. The draft's lines add up to less than the total it shows, because line 9 is short by 8 units. A good app flags the mismatch between the line sum and the printed total. The clerk catches it in two minutes instead of finding a stock difference at month-end.
Then the old rule still applies: before this bill is paid, it must match the purchase order and the goods actually received (chapter 2.2). AI read the bill correctly or not — it cannot tell you whether 48 units really came off the truck. Only the GRN can.
Documents that work well, and badly
| Reads well | Needs extra care |
|---|---|
| Printed company invoices | Handwritten orders from the beat |
| Typed price lists | Carbon copies, faded thermal slips |
| Bank slips and UPI screenshots | Photos at an angle or with fingers over numbers |
| Clear credit notes | Bills with corrections written over the print |
Common mistake
Throwing away the paper, or deleting the photo, once the entry is posted. Months later the company disputes a claim, or the auditor asks why the stock of one product is off, and the only evidence is an entry the AI typed. Every entry starts with a document (chapter 1.2) — keep the document attached to the entry it produced.
Where you'll use this
This automates the oldest rule in the book: the document is the evidence and the entry comes from it (chapter 1.2), and the bill must match the order and the goods received (chapter 2.2). AI only removes the typing between paper and entry. How to check its draft properly is chapter 4.17.
Check yourself
- What is the difference between a proposed entry and a posted one?
- Why did the draft's total match the paper even though line 9 was wrong?
- AI read a supplier's bill perfectly. Can you pay it now?
Answers
- A proposed entry is a draft that has not touched the books; it becomes posted only when a person approves it.
- It copied the printed total instead of adding up its own lines, so the lines and the total disagreed.
- Not yet. It must still match the purchase order and the goods actually received (the GRN).